IGWO-RF: A Hybrid Improved Gray Wolf Optimizer and Random Forest Wrapper for High-Dimensional Feature Selection

Abstract

Not all data features are crucial for uncovering hidden knowledge within various datasets, making feature selection a significant area of interest. This work proposes IGWO-RF, a new meta-heuristic algorithm that combines an improved gray wolf optimization (GWO) algorithm with random forest (RF) for feature selection. The improved GWO introduces a nonlinear convergence parameter for better exploration- exploitation balance and a GA-inspired crossover operation using alpha and beta wolves to accelerate convergence. The RF algorithm evaluates the fitness of feature subsets in each iteration. The proposed technique was evaluated on 10 benchmark UCI datasets (including Wine, Sonar, Vehicle, and Parkinson's) based on the average number of selected features, average classification accuracy, and best fitness. Comparative analysis with four popular wrapper-based methods (GWO-RF, ACO, PSO, ABC) demonstrated the superiority of IGWO-RF. Specifically, IGWO-RF achieved the highest average classification accuracy of 91.23% using the SVM classifier, outperforming GWO-RF (89.91%), PSO (89.12%), and ABC (87.94%). Furthermore, IGWO-RF obtained the most compact feature subsets, selecting on average only 31.22% of the original features in the Glass dataset and 20.89% in the Vehicle dataset—a significant reduction compared to other methods. The algorithm also showed faster convergence and reduced execution time. Therefore, IGWO-RF proves to be an effective approach for enhancing pattern classification performance through efficient feature selection.